Beyond Ghost Work: Centering Disabled Expertise in the Age of AI

Disabled Experts Demand Meaningful 'Human-in-the-Loop' Agency Over AI Bias.

The narrative surrounding artificial intelligence is frequently sanitized, projecting an idealized vision of seamless automation, frictionless efficiency, and objective technological progress. Yet, as the critical analysis of Aranya Sahay’s film Humans in the Loop demonstrates, this pristine illusion is entirely maintained by the hidden, precarious labor of marginalized workers who are tasked with standardizing a complex world into rigid data points. The struggle of Nehma, an Adivasi woman forced to suppress her deep contextual and ecological knowledge to label caterpillars merely as agricultural pests, is not an isolated incident of corporate oversight. Rather, it exposes the foundational logic of contemporary machine learning: a systemic devaluation of nuanced, localized human knowledge in favor of standardized, corporate-mandated categorizations.

When we extend this critical lens into the realm of disability studies, a profound and troubling parallel emerges. The identical mechanisms that exploit invisible labor in the data labeling industry are the exact mechanisms through which artificial intelligence systems exclude, misinterpret, and actively harm disabled communities. Both paradigms are inherently reliant on an assumed "normal" baseline—a standard user, a typical body, a conventional way of navigating and processing the world. Consequently, both systems treat any deviation from this imagined normality as a structural problem that must be solved, corrected, or simply excluded from the dataset entirely. To disrupt this pervasive cycle of technoableism, we must radically reimagine the widely touted concept of the "Human in the Loop" (HITL). It is imperative that we move beyond the extraction of marginalized labor for data annotation and instead position people with disabilities as the essential, guiding experts at every stage of the AI lifecycle.

The Imperative of Human Oversight in an Ableist Architecture

Artificial intelligence is rapidly expanding its footprint across critical sectors, deeply integrating into healthcare diagnostics, educational frameworks, employment screening, and independent living technologies. For individuals with disabilities, these emerging technologies hold the potential to offer profound, life-altering opportunities. Innovations such as image description generation, environmental control systems, accessible text formatting, and personalized learning navigation can significantly augment autonomy. However, the societal stakes associated with these deployments are exceptionally high. When AI systems inevitably make errors, the consequences for disabled individuals transcend mere technological inconvenience. These failures can severely restrict access to essential services, reinforce historical patterns of discrimination, reduce hard-won independence, and even place individuals at tangible physical or social risk.

This risk is not anomalous; it is baked into the very architecture of how AI operates. Machine learning models function essentially as massive pattern-matching engines; they excel at identifying statistical correlations within historical training data, but they completely lack the capacity to comprehend the social, cultural, or personal contexts that generate those patterns. Disability is a lived experience that is inherently complex, dynamic, and entirely defiant of the simplistic categorizations that algorithms require to function. Two individuals sharing the exact same medical diagnosis may possess vastly divergent abilities, preferred communication methods, and daily support requirements. Furthermore, a single individual's access needs are rarely static; they can fluctuate dramatically based on environmental factors, varying levels of fatigue, chronic pain, or systemic stress.

When artificial intelligence systems encounter this dynamic human diversity without appropriate oversight, they frequently interpret difference as a critical error or a risk factor. For example, facial recognition and analysis systems may completely misread the expressions of an individual with facial paralysis. Behavior-monitoring software deployed in educational or workplace settings might flag autistic stimming or movement as non-compliance, inattention, or suspicious behavior. Similarly, standard speech recognition software routinely fails to process or accurately transcribe dysarthric speech, effectively silencing users. In healthcare, diagnostic algorithms may underestimate the quality of life of a disabled person simply because the underlying training data is polluted with negative, ableist assumptions about living with a disability. Without an expert human in the loop to provide necessary context, question the algorithm's biased assumptions, and recognize when an output completely fails to reflect the user's actual circumstances, AI simply operates as an engine for automated, scalable ableism. However, this oversight is only effective if the human reviewer is properly trained and free from the very biases embedded within the machine.

Redefining the Expert: From Marginalized Subjects to Central Authorities

Historically, the technology industry has approached the human in the loop concept with a remarkably narrow and self-serving lens. It is often operationalized as a superficial oversight mechanism, where a lower-level staff member simply signs off on an algorithm's decision without meaningful scrutiny. Alternatively, on the rare occasions when disabled people are included in the development pipeline, they are almost exclusively relegated to the role of passive beta-testers at the very end of the cycle, asked merely to interact with a finished product. This reactive, afterthought approach is fundamentally incompatible with the creation of genuinely inclusive technology.

A truly transformative approach to artificial intelligence requires a paradigm shift: acknowledging that disabled people are not just passive beneficiaries, end-users, or subjects of technological intervention. They are the definitive experts in their own lives, possessing unparalleled insight into their access requirements and the systemic mechanisms of exclusion they navigate daily. This specialized expertise is invaluable and must be embedded at the very genesis of technological development. True inclusion demands that disabled individuals lead the processes of identifying the core problem an AI intends to solve, defining the metrics of what constitutes a successful outcome, and anticipating potential harms before a single line of code is written.

Consider the widespread deployment of AI-powered recruitment and applicant tracking tools. To a non-disabled developer or a corporate human resources department, such a system might appear as a marvel of efficiency, capable of sorting through thousands of resumes in seconds to find the "ideal" candidate. However, disabled experts integrated into the design process can immediately identify how such a system might unfairly penalize applicants. They can point out that algorithms often downgrade candidates who possess employment gaps due to necessary medical care, who have non-standard career trajectories, who utilize alternative communication styles, or whose application formatting reflects the use of specific assistive technologies. Without disabled people acting as the authoritative human in the loop to expose these hidden, discriminatory patterns, these algorithms will continue to silently gatekeep opportunities under the deceptive guise of objective efficiency.

Agency, Support, and Resisting Automation Bias

The integration of disabled expertise also directly addresses the critical, delicate balance between providing technological support and usurping human control. Artificial intelligence can be highly effective in organizing vast amounts of information, offering predictive text, suggesting communication options, or streamlining exhausting repetitive tasks. However, it is an absolute imperative that this support does not seamlessly morph into control over the user.

For individuals who utilize augmentative and alternative communication (AAC) devices, AI-driven word prediction can drastically accelerate typing speeds, which is particularly vital for those who use eye gaze technology, switch access, or other alternative controls. Yet, it is vital that the system never assumes the authority to autonomously speak on the user's behalf without explicit approval. An AI might rapidly generate a suggested sentence that is grammatically flawless, but entirely misses the user's intended emotional tone, specific sense of humor, or personal identity. The disabled user must permanently retain the absolute authority to accept, edit, reject, or completely ignore the machine's suggestions. The fundamental goal is not merely to have a professional supervising the technology, but to guarantee that the disabled individual retains total ownership and authority over decisions affecting their own voice and life.

Maintaining this agency requires a constant, vigilant organizational resistance against "automation bias"—the dangerous psychological tendency to trust a computer-generated recommendation simply because it presents itself as scientific, objective, or data-driven. In high-stakes domains such as healthcare assessments, social protection eligibility, educational interventions, and access to public services, staff may gradually begin to defer to the AI's judgment to cut costs or save time. When this happens, human oversight devolves into a meaningless formality, rubber-stamping the machine's biases. To actively combat this, organizations must establish clear, non-negotiable thresholds detailing exactly when an automated decision requires mandatory, rigorous human review. This includes scenarios where the system registers low confidence, where the disabled user disputes the outcome, where the consequences of the decision are severe, or where specific disability-related factors are likely to skew the algorithm's accuracy. The purpose of human involvement is not to provide indefinite administrative cover for a flawed system, but to continuously generate critical evidence that forces structural improvements.

Building an Accessible Loop and Ensuring Accountability

It is a profound paradox to advocate for disabled people as the essential experts in the loop if the loop itself is structurally inaccessible. Oversight mechanisms, appeals processes, and feedback channels are entirely performative if they cannot be readily navigated by the very marginalized populations they are theoretically designed to protect. If an AI system makes a decision that negatively impacts a disabled person's life, that individual must have the absolute right to know that an algorithmic process was involved. They must possess the right to demand a transparent explanation, the right to correct inaccurate underlying data, and the right to appeal the decision directly to a human being who holds the actual authority to overturn the algorithmic outcome.

Crucially, this appeals process must be universally accessible. An oversight mechanism is not genuinely inclusive if it relies exclusively on complex, jargon-heavy written forms, visually demanding web interfaces, or telephony systems that immediately exclude individuals with communication or hearing disabilities. True inclusion requires proactively providing Easy Read formats, sign language interpretation, accurate captioning, robust advocacy support, text-based communication alternatives, and extended timeframes for response. A human in the loop is only a meaningful, protective safeguard if the affected individual can actually reach that human and be fully heard.

Ultimately, the conversation surrounding the human in the loop is fundamentally a conversation about power and accountability. When an AI system inflicts harm or perpetuates discrimination, the ethical and legal responsibility cannot be deflected onto an abstract algorithm or a black-box neural network. The developers who code the systems, the procurement teams who purchase them, and the public authorities or private organizations who deploy them bear the ultimate, irreducible responsibility for their societal impact.

This accountability requires viewing co-design not as a one-off, unpaid consultation, but as a continuous, iterative lifecycle. AI systems are dynamic; they evolve as datasets grow and user behaviors shift. A technological tool that tested as perfectly accessible during a highly controlled pilot program might easily erect entirely new barriers when deployed into complex, real-world settings. Consequently, disabled people must be fairly and consistently compensated for their ongoing expertise in evaluating and refining these systems post-deployment. Furthermore, this participation must be radically diverse, capturing the wide spectrum of physical, sensory, cognitive, communication, and psychosocial disabilities, rather than relying on a single, monolithic representation of the disabled experience.

Core Principles for the Expert Human in the Loop

To operationalize the vital inclusion of disabled experts in AI oversight and development, organizations and technology developers must commit to the following foundational principles:

• Context Over Categorization: AI development must relentlessly prioritize the lived, contextual realities of disabled people over the rigid, standardized metrics that algorithms typically favor. Human review must actively protect the nuance and localized knowledge that data labeling processes so frequently destroy.

• Agency Amplification: Artificial intelligence should be utilized exclusively as an instrument to strengthen human agency, relationships, expertise, and choice. It must never be permitted to override the individual autonomy of disabled users under the pretext of operational efficiency or predictive accuracy.

• Continuous, Compensated Co-Design: The involvement of disabled people cannot be relegated to post-design beta testing. It must be an ongoing, financially compensated collaboration that spans initial conception, data selection, model training, and continuous post-deployment auditing.

• Irreducible Human Accountability: Algorithms do not hold ethical responsibility; human institutions do. There must always be a clear, highly accessible pathway to a human decision-maker who possesses the institutional authority to halt a system or immediately reverse an automated harm.

The Inclusive AI Audit Checklist

For developers, organizations, and policy frameworks seeking to implement a robust, disability-centered Human in the Loop, this checklist provides a starting metric for systemic accountability:

• Pre-Development & Design:

  • Have disabled individuals from diverse backgrounds been compensated to help define the core problem this AI is attempting to solve?

  • Does the training data respect the nuanced, contextual realities of the target population, or does it enforce ableist categorizations that treat deviation as an error?

  • Is the system intentionally designed to suggest options and support the user, rather than autonomously executing final decisions on their behalf?

• Deployment & Oversight Mechanisms:

  • Are the individuals acting as the human in the loop during the operational phase explicitly trained to recognize and reject ableist bias?

  • Is there a strict, documented protocol that triggers mandatory human review for decisions involving high stakes (e.g., healthcare access, employment screening, social protection)?

  • Are organizations actively monitoring for and counteracting "automation bias" among the staff responsible for system oversight?

• Accessibility of Recourse:

  • Can a user easily and transparently discover when an AI system has influenced a decision regarding their access, rights, or eligibility?

  • Is the mechanism to challenge or appeal the AI’s decision available in multiple accessible formats (e.g., Easy Read, sign language, text-based alternatives, extended timeframes)?

  • Does the appeals process connect directly to a human with the institutional authority to immediately overturn the algorithm's recommendation?

The future of artificial intelligence does not lie in the complete, frictionless automation of human existence, nor does it lie in the continued extraction of invisible labor to build vast systems of exclusion. It lies in recognizing that the "glitches" algorithms encounter when processing disability are not errors residing in the human body, but profound failures in the machine's design and the ableist assumptions of its creators. By placing disabled experts firmly and authoritatively within the loop, we can build technology that does not merely categorize the world, but actually understands and respects the full spectrum of human diversity.

Sources

• Banes, D. (2026). Human in the Loop: Why Inclusive AI Must Keep People with Disabilities at the Centre. Medium. https://davebanesaccess.medium.com/human-in-the-loop-why-inclusive-ai-must-keep-people-with-disabilities-at-the-centre-1d6bbdfbca82
• Singit, N. (2025). Human in the Loop: Artificial Intelligence, Disability, and Hidden Ableism. The Bias Pipeline. https://thebiaspipeline.nileshsingit.org/2025/12/human-in-loop-artificial-intelligence.html
• Sahay, A. (Director). (2024). Humans in the Loop [Motion Picture].
• Mehrotra, K. (2022). Human Touch: The invisible army of workers training artificial intelligence across India. FiftyTwo. https://fiftytwo.in/story/human-touch/
• Foundational & Theoretical Frameworks
• Gray, M. L., & Suri, S. (2019). Ghost Work: How to Stop Silicon Valley from Building a New Global Underclass. Houghton Mifflin Harcourt.
• Shew, A. (2023). Against Technoableism: Rethinking Who Needs Improvement. W. W. Norton & Company.
• Costanza-Chock, S. (2020). Design Justice: Community-Led Practices to Build the Worlds We Need. MIT Press.
• Whittaker, M., Alper, M., Bennett, C. L., Hendren, S., Kaziunas, E., JafariNaimi, N., Burl, M., & West, S. M. (2019). Disability, Bias, and AI. AI Now Institute. https://ainowinstitute.org/disabilitybiasai-2019.pdf
• Parasuraman, R., & Manzey, D. H. (2010). Complacency and bias in human use of automation: An attentional integration. Human Factors: The Journal of the Human Factors and Ergonomics Society, 52(3), 381–410.